#156 - OpenAI's Sora, Gemini 1.5, BioMistral, V-JEPA, AI Task Force, Fun!

Last week saw significant advancements in the field of AI, specifically in the areas of text-to-video models and policy discussions. OpenAI's new text-to-video model, Sora, stole the spotlight with its impressive capabilities, showcasing high-resolution, clear videos generated from text inputs. Although some may argue that Gemini 1.9 was also a notable development, the general consensus seems divided. Sora's technology, which uses a transformer model to embed meaning from images, marks a significant leap forward in text-to-video AI. Despite the excitement, it's important to note that OpenAI did not release the specifics of their model's setup in their report. On the policy front, Jeremy Harris, one of the hosts, made an announcement about a potential career opportunity for an experienced AI team leader. The industry veteran, Ben, has experience working at a Fortune 100 company and holds a PhD in physics. Interested parties can reach out to Gladstone AI for more information. Lastly, the hosts introduced a new segment called the "fun section," where they'll share less serious stories that don't quite fit into the other sections. They closed the episode with this new addition, aiming to add some levity to the discussion. Overall, the past week demonstrated the rapid progress being made in AI, from impressive technological breakthroughs to important policy discussions. Stay tuned for more updates on Last Week in AI.

OpenAI has developed a new model called Sora, which can generate meaningful and physically accurate videos based on text descriptions. This model goes beyond language models by also incorporating video and text data, creating what OpenAI calls "spacetime patches" – atomic units of meaning in the context of video generation. These spacetime patches enable the model to learn and understand physical world models, including laws of physics, as evidenced by its ability to accurately portray real-world phenomena like glasses shattering or balls falling. Sora is not just a language model; it's a diffusion transformer, which generates videos end-to-end, and it can perform various tasks like text-to-video, image-to-video, and video-to-video editing. The impressive results are a significant breakthrough in the field of AI-generated videos and demonstrate OpenAI's ongoing belief in the robustness of transformers to learn world models. However, Sora is currently only available to a select group of red teamers, artists, designers, and filmmakers for assessment and feedback.

OpenAI's latest model, Sora, has demonstrated the ability to understand and track objects over long time horizons, showcasing an emergent understanding of object permanence. This breakthrough aligns with OpenAI's mission to build AGI and their consistent theme of chunking up data and looking at it from the right perspective before applying massive scale. Google's Gemini 1.5 Pro, a mid-sized multimodal model, was also announced, boasting a large context window and reportedly being as good as, if not more efficient than, their higher-tier model. This new model, with its impressive context window and potential competitive pricing, could add pressure on OpenAI. Sora's object permanence understanding and OpenAI's data chunking approach are significant advancements, offering a glimpse into the future of AI research.

A new AI model has been developed which can process and recall information from context windows that are over a million tokens (approximately 750,000 words) long. This is a significant improvement over previous models, as they often forgot details mentioned early in their prompts when dealing with such large context windows. This new model, which is incredibly powerful, achieves near-perfect recall in needle and haystack tests, outperforming all previous models including GPT-4 and Gemini Ultra. The exact mechanism by which this recall is achieved is not yet clear, but it may involve some form of stateful memory or algorithmic modification. This breakthrough allows the system to learn and absorb context incredibly quickly, even picking up obscure languages like Kalamang, which has fewer than 200 speakers worldwide. It's likely that this improvement has been achieved through tweaks to the decoding and prompting processes, but further research is needed to understand the full extent of this impressive advancement in AI technology.

The Gemini 1.5 Pro model outperforms expectations in making qualitatively better predictions than what was previously thought possible, even surpassing the power law fit. This could be due to a fundamental algorithmic improvement or an unexpected scaling effect. On a different note, GrogQ Inc's new language processing unit, GrogQ, has gained attention for its impressive 500 tokens per second throughput, which is four times faster than other inference services. However, its limited onboard RAM might limit the number of customers it can serve at a given time. These advancements highlight the ongoing innovation in AI and language models.

The new language processing unit (LPU) chip, which is specifically designed for inference in large language models (LLMs), is a significant breakthrough in the field of AI. However, it also comes with limitations, such as the high cost and the fact that it only does inference and not training. This means that companies will need to invest heavily in infrastructure to use this chip effectively. Furthermore, the trend towards models doing more of their thinking during inference rather than training suggests that we can expect to see more specialized chips for LLM use cases in the future. Another interesting development is Scenario's new tool that allows video game developers to create consistent assets of a character from a single reference image. This is a significant step forward in addressing the challenge of creating consistent character generation, which has been a long-standing issue in industries like animation, webcomics, and video games. The tool, which generates IP for commercial applications, raises questions about copyright and IP protections. OpenAI, which already has a web crawler (gpt bot), is also reportedly working on a web search product using its language model technology. This is not surprising given Microsoft's use of gpt for its customized search product and OpenAI's partnership with Microsoft. Overall, these developments highlight the rapid pace of innovation in the field of AI and the increasing importance of specialized hardware for inference in LLMs.

The AI landscape is becoming increasingly crowded as more companies, including Google and Adobe, introduce AI-enabled products. The latest addition is Adobe Acrobat's AI Assistant, which allows users to interact with documents in a chat-like manner. Meanwhile, OpenAI, a leading player in the AI space, has a peculiar ownership structure for its venture capital fund, with Sam Altman personally owning it despite OpenAI's involvement in its operations. This arrangement raises questions about governance and potential risks, especially considering the significant investments made by the fund. The market share imbalance between Google and OpenAI, along with competition from other players, adds to the uncertainty of OpenAI's ability to make a substantial impact. However, the rapid advancement of AI technology and the growing demand for AI-integrated tools suggest that the competition will only intensify.

The intersection of technology and business continues to evolve rapidly, with companies like OpenAI, Reddit, and NVIDIA pushing the boundaries of what's possible in the realm of AI. OpenAI, an organization known for its advanced AI research, has been keeping a secret fund for lawyers, adding to the intrigue surrounding the organization's unusual structure. Reddit, meanwhile, has signed a significant AI content licensing deal, reportedly worth $60 million annually, which could set a precedent for future deals. And NVIDIA has revealed its EOS supercomputer, a massive AI processing machine with 4,608 H100 GPUs, marking a notable achievement in the world of individual supercomputers for AI applications. These developments underscore the growing importance of AI and the significant investments being made in this area. The future is sure to bring more advancements and innovations as these companies and others continue to push the boundaries of what's possible.

The advancements in AI technology are progressing at an unprecedented pace. EOS, a supercomputer, was able to train a large language model on one billion tokens in just four minutes, which was previously a three and a half year long process. Google also launched an internal AI model named Goose, which is aimed at helping employees write code faster by utilizing the internal text tags of the company. This model is reportedly trained on 25 years of engineering expertise at Google. Chinese startup Moonshot AI raised a billion dollars in funding for its open AI play, having launched a smart chat bot, KimiChat, in October. Moonshot AI's large language model, moonshot LLM, is capable of processing up to 200,000 Chinese characters in its context window. These advancements demonstrate the significant investments being made in AI technology and the rapid progress being made in the field. Additionally, it's noteworthy that these developments involve collaboration between various parts of tech giants like Google and the involvement of military-affiliated institutions in China. These context windows, while impressive, should be evaluated not just by their size, but also by the quality of the AI's handling of the context.

The AI industry, particularly in the Chinese market, is seeing significant investment and innovation. Chinese models are making headlines with large fundraises, but a cautious approach is recommended due to the potential focus on vanity metrics. Lambda, a competing firm, raised 320 million and specializes in compute and generative AI training, making them a notable player. Another company, ex sales, raised 110 million for AI agents for businesses, which are enabling various interactions between businesses and customers. Open source developments include Bio Mistral, a collection of pre-trained large language models for medical domains, which outperforms other models on various medical question answering tasks. The model was trained using the French National Center for Scientific Research's high-performance computer, demonstrating the impact of giving researchers access to powerful hardware. NOMIC AI released the first fully open source long context text embedding model, Norney Symbed text V1, which surpasses OpenAI's performance on various benchmarks. This model can handle sequence lengths of 8,000 tokens and is available under an Apache 2 license. These developments showcase the growing importance and potential of AI in various industries and the ongoing advancements in the field. The open-source nature of many of these projects allows for collaboration and further innovation, leading to more advanced models and applications.

Researchers are continuously pushing the boundaries of artificial intelligence (AI) by developing advanced models and techniques to improve training and understanding of various forms of data, such as text and video. The first example discussed a paper on a new text model called "Longformer," which uses long context windows and a small model size to achieve superior performance on short and long context benchmarks, surpassing OpenAI's text embedding models. The researchers emphasized the importance of having a small, efficient model that can effectively process long contexts, which was a challenge until now. They also noted the openness and accessibility of the project, making the code, data, and everything else readily available to the public. The second story featured Meta's new V-Jepa AI model, which aims to predict patches of video based on the idea that closely related information often appears in neighboring patches or frames. This model, like OpenAI's Sora, operates on the embedding space and takes advantage of the fact that meaning is similar in closely related parts of the video. The ultimate goal is to train an encoder that can extract meaningful embeddings from patches of video or images. Both projects demonstrate the ongoing efforts to create more advanced and versatile AI models, with a focus on making the most of available resources and pushing the boundaries of what's possible in the realm of text and video understanding.

The research presented in the discussed papers showcases different approaches to achieving artificial general intelligence (AGI) and the importance of both scalability and specialized architectures. The first paper introduces VJAPA, a model for unsupervised learning from video, which has limitations compared to more advanced models like Sora, but reflects a commitment to Jan LeCun's vision of AGI and explores feature prediction as a new objective for unsupervised learning. The model is primarily a research effort, focusing on cell supervised training and can handle large amounts of unlabeled video data, which could lead to more efficient scaling. The second paper discusses the ability of language models to perform chain of thought reasoning without explicit prompting. The researchers show that using the top K alternative tokens during decoding can reveal the chain of thought paths inherent in the sequences, potentially increasing confidence in the final answer when a chain of thought type output is present. This finding highlights the inherent capabilities of language models and could lead to new applications. Both papers demonstrate the ongoing research in the field of AGI and the importance of exploring various approaches, from specialized architectures to scalability, to make progress towards human-level intelligence. The findings in these papers could potentially lead to more efficient methods for training large models and unlocking new capabilities in AI systems.

Recent advancements in AI research, specifically in the areas of longer context windows and emergent abilities, are pushing the boundaries of what's possible in the field. Firstly, the development of ring attention mechanisms in transformer models allows for longer context windows, theoretically with no limit, by passing intermediary keys and values between multiple devices in a ring-like structure. This is a significant step towards achieving more comprehensive understanding and context in AI. Secondly, Amazon's AGI team has made strides in text-to-speech technology with their model, Big Adaptive Streamable TTS, which has shown emergent abilities, handling complex aspects of text, such as foreign words and punctuation, without explicit training. Moreover, the collaboration between OpenAI and Microsoft to document the use of their systems by foreign hackers highlights the importance of safety and security in AI research and application. Hackers have been using AI for relatively mundane tasks, such as drafting emails and debugging code, and OpenAI and Microsoft are working to prevent such unintended uses. These developments underscore the ongoing trend towards longer context and more sophisticated capabilities in AI, with researchers and companies continually pushing the envelope to unlock new possibilities. It's an exciting time for AI research and development, with numerous potential applications and implications for various industries and society as a whole.

State-affiliated hacking groups, such as Charcoal Typhoon (Chinese), Salmon Typhoon (Chinese), and Forest Blizzard (Russian), are increasingly using advanced technologies, including OpenAI models, to translate technical papers and gain insights into intelligence agencies and potential targets. These groups have a broad operational scope, targeting sectors like government, higher education, communications, oil and gas, and have a history of targeting U.S. defense contractors and cryptographic tech companies. The Russian group Forest Blizzard has been linked to GRU unit 26165 and has been active in the context of Ukraine. In response, the U.S. House of Representatives has launched a bipartisan AI task force, led by Speaker Mike Johnson and Minority Leader Hakeem Jeffries, to support AI innovation and study potential threats. The task force will have 24 members and will be chaired by J. Obernolte, a computer science expert and video game developer, and co-chaired by Ted Lieu, a computer science background and known hawk on AI issues. The task force aims to canvas around for a wide range of opinions on where the field might be going and account for the fact that a lot is unknown, especially regarding the potential speed of technological advancements.

Our digital interactions, even seemingly insignificant ones like swiping on a touchscreen, could potentially reveal sensitive information such as fingerprints. Researchers from the University of Colorado and China have developed a side channel attack that can reproduce about 28% of partial fingerprints and 10% of complete fingerprints based on the sounds made while swiping. This highlights the ease with which information can be gathered from our environments and the increasing importance of monitoring and safeguarding data. Another significant development is the increasing number of AI-related bills being introduced in the US, with 15 new bills per week and a total of 407 bills across more than 40 states. New York, California, Tennessee, and Illinois are among the states leading the charge. In the realm of AI ethics, a ruling in Canada set a legal precedent for AI alignment, with Air Canada being found liable for inaccurate information provided by its chatbot. The decision implies a legal requirement for reasonable care in ensuring AI systems give true outputs, opening up interesting discussions about the responsibilities and accountabilities of AI developers and users. Lastly, the FTC issued a warning about potential quiet changes to Terms of Service (ToS) for AI training data, with companies potentially altering their privacy policies to use user data without restriction. The FTC encourages users to stay informed and vigilant about their data and privacy.

Companies must be transparent about their privacy policies and cannot secretly change them to make money from user data without consent. This was highlighted in the FTC's recent warning to firms, including Zoom, which updated its terms of service in August 2023 to allow the use of user data for AI training without an opt-out option. The lawsuit against OpenAI by Sara Silverman and others was partially dismissed in a California court, leaving only the claims of direct copyright infringement and unfair competition. The judge dismissed the other claims due to a lack of clear economic injury and the speculative nature of the risk of future damage to intellectual property. These rulings could set a precedent for limiting the types of lawsuits that can be brought against LLM companies. The visual guide to Mamba and state space models by Martin Ruten Dorst provides a detailed explanation of the architecture and concepts behind these models, which are built on control theory and hardware optimization. These topics can be quite technical, but the guide offers a comprehensive understanding of the subject matter.

The use of artificial intelligence (AI) in scientific research and creative industries is a topic of growing concern. The discussion highlighted two instances where AI was involved in producing misleading or false information. The first instance involved a retracted scientific paper that contained AI-generated images, and the second instance involved an AI-generated speech read at an awards ceremony by Helen Mirren. Both incidents raised questions about the reliability and authenticity of AI-generated content and sparked a backlash against the use of AI in these fields. The Microsoft Super Bowl ad, which aimed to reassure people that AI is here to help make their dreams come true, seemed to be an attempt to reframe the narrative around AI in a more positive light. However, the incidents discussed serve as a reminder that there is a need for greater oversight and transparency in the use of AI in research and creative industries to prevent the spread of misinformation and maintain the integrity of these fields.

Microsoft's recent Super Bowl ad showcases their efforts to reframe Artificial Intelligence (AI) as an enabler of human achievement rather than a replacement or source of fear. The ad, which features Microsoft's co-pilot AI bot, aims to challenge the cultural perception of AI as a risk or threat, much like Apple did in their 1984 Super Bowl ad. Although some viewers found the ad to be underwhelming, its message is clear: AI exists to help us, not replace us. Microsoft's marketing strategy is a reflection of the ongoing conversation about the role of AI in our lives and the potential it holds for human advancement.